Learning/ AI Healthcare Quality & Safety/ Lesson 01
Lesson 01 of 10AI Healthcare Quality & Safety

What Is AI?
From Algorithms to Clinical Intelligence

Artificial intelligence is reshaping healthcare faster than any technology since the electronic health record. Before applying it, leading it, or governing it, every healthcare professional needs a clear, honest understanding of what AI actually is — and what it is not.

What you will learn
Define artificial intelligence and distinguish it from related terms including machine learning and deep learning
Explain why healthcare is both uniquely suited to and uniquely challenged by AI adoption
Describe the difference between narrow AI and general AI and why this distinction matters clinically
Identify the primary categories of AI application in healthcare settings
Articulate the limitations of AI that every clinical professional must understand
Start Lesson Back to Course
Lesson Snapshot
Lesson01 of 10
Progress10% Complete
Est. Time~35 Minutes
Knowledge Checks5 Questions
Course Progress
01 What is AI? 02 How AI Learns 03 Computer Vision & Imaging 04 Predictive Analytics 05 NLP & Documentation 06 Clinical Decision Support 07 Safety & Risk 08 Bias & Equity 09 AI Governance 10 Leading AI-Ready Orgs

Defining artificial intelligence
clearly and honestly

Artificial intelligence is a broad term describing computational systems designed to perform tasks that would typically require human intelligence — tasks like recognizing patterns, making predictions, interpreting language, and generating recommendations. It is not a single technology. It is a family of approaches, each with different capabilities, limitations, and appropriate applications.

The most important distinction for healthcare professionals is between narrow AI and general AI. Narrow AI — sometimes called weak AI — is the only form that currently exists in clinical practice. A narrow AI system is designed to perform one specific task with high proficiency: reading a chest X-ray, predicting a patient's risk of sepsis, or flagging a potential drug interaction. It cannot generalize beyond its trained task. It cannot reason about context it was not trained to recognize. It cannot adapt to situations outside its design parameters.

General AI — a system that can reason across domains, learn from context, and perform any intellectual task a human can — does not exist. It remains a theoretical concept. When clinicians or administrators encounter claims about AI that sound unlimited or universal, those claims deserve skepticism.

Critical Distinction

Every AI system in clinical use today is narrow AI — designed for one task, trained on specific data, and limited to the context of its training. No AI system in healthcare can reason, generalize, or judge the way a clinician can.

Why healthcare and AI
are both natural partners and difficult ones

Healthcare generates more data per person than almost any other domain — clinical notes, laboratory results, imaging studies, vital signs, medication records, genomic data, and increasingly patient-generated data from wearables and remote monitoring devices. AI systems learn from data. The richness of healthcare data creates enormous potential for AI to find patterns, identify risks, and support decisions at a scale and speed that humans cannot match alone.

At the same time, healthcare presents challenges that make AI adoption more complex than in other industries. Clinical data is noisy, incomplete, inconsistently structured, and heavily influenced by documentation practices that vary between clinicians, institutions, and countries. Patients are diverse in ways that training datasets frequently underrepresent. The consequences of error are measured in human harm — not conversion rates or ad clicks. And the relationship between clinical judgment and algorithmic output is mediated by human factors that AI cannot fully account for.

These characteristics do not make AI inappropriate for healthcare. They make careful, governed, transparent adoption essential — which is why the skills this course develops matter.

The primary categories
of AI in clinical practice

AI in healthcare falls into several broad application categories, each with distinct characteristics and governance considerations. Diagnostic AI systems analyze clinical data — imaging, laboratory values, pathology slides, or clinical notes — to detect, classify, or characterize disease. These systems assist or augment diagnostic decision-making and include radiology AI tools, AI-powered pathology analysis, and AI-enabled dermatology screening.

Predictive AI systems analyze patient data to forecast future clinical events — deterioration, readmission, sepsis onset, length of stay, or mortality risk. These systems inform proactive clinical decisions and early intervention. Clinical decision support AI integrates with electronic health records to provide real-time recommendations, alerts, and reminders at the point of care. Natural language processing systems analyze unstructured clinical text — notes, dictations, discharge summaries — to extract structured information, automate documentation, or support coding and billing accuracy.

Administrative and operational AI automates or optimizes non-clinical processes — scheduling, resource allocation, supply chain, and revenue cycle management. Understanding which category a specific AI system belongs to is the first step in assessing its governance requirements and clinical risk profile.

Key Principle

The category of AI — diagnostic, predictive, decision support, NLP, or operational — determines its clinical risk level, governance requirements, and the oversight model that should surround it. Category is the first question in any AI governance assessment.

Key concepts
from this lesson

Key Concept

Artificial Intelligence

Computational systems designed to perform tasks requiring human-like intelligence — pattern recognition, prediction, language interpretation.

Key Concept

Narrow AI

AI designed for one specific task — the only form of AI that currently exists in clinical practice.

Key Concept

Machine Learning

A subset of AI where systems learn patterns from data rather than following explicitly programmed rules.

Key Concept

Deep Learning

A subset of machine learning using neural networks with many layers — powers most modern imaging and NLP AI applications.

Key Concept

Training Data

The dataset on which an AI model learns — the quality, diversity, and representativeness of training data determines model performance.

Key Concept

Generalization

The ability of an AI model to perform on new data beyond its training set — a key limitation of narrow AI systems.

Case Study

The AI system that couldn't see what it hadn't learned

A regional hospital implements an AI-powered sepsis prediction tool trained on data from academic medical centers in the United States. The model achieves a reported sensitivity of 82% and specificity of 91% in its validation study — strong performance by most clinical standards.

Twelve months after deployment, the hospital's patient safety team notices that the sepsis prediction model performs significantly worse for patients admitted from long-term care facilities — a population that represents 28% of the hospital's admissions but only 4% of the training dataset.

The model had never encountered enough examples of long-term care patients — whose baseline laboratory values, documentation patterns, and comorbidity profiles differ systematically from the academic center population — to learn to recognize their deterioration signals reliably. It was not malfunctioning. It was doing exactly what it had been trained to do. The training data simply did not represent the population to whom it was being applied.

The hospital suspends the tool for long-term care admissions, commissions a bias assessment, and begins the process of retraining with locally representative data — a process that takes seven months.

What this illustrates

AI systems do not generalize beyond their training. Deploying a model in a population it was not trained on is one of the most common — and most preventable — causes of AI-related clinical harm. Understanding what an AI was trained on is not optional due diligence. It is foundational governance.

Reflection Prompt

What AI systems are already in your organization?

Before this course progresses to the technical and governance dimensions of healthcare AI, take time to inventory what AI systems currently influence clinical or operational decisions in your organization. You may be surprised by how many exist — some are obvious, others are embedded in systems you use daily. Which category does each belong to? What do you know about what data each was trained on? And who in your organization is accountable for governing them?

↗
Further Learning

IHI's work on digital health and AI safety is available through their publications and white papers at ihi.org. The IHI Open School Quality Improvement series also provides foundational context for understanding technology-enabled improvement in healthcare systems.

Knowledge Check — Lesson 01

1. A hospital implements an AI system that analyzes echocardiogram images to detect left ventricular dysfunction. This system is best classified as:

AGeneral AI — because it performs a complex clinical task
BNarrow AI — designed for one specific diagnostic task
CDeep learning only — not AI in the clinical sense
DClinical decision support — because it supports physician decisions

2. Which of the following most accurately describes why healthcare data is particularly challenging for AI training?

AHealthcare data is too small in volume for AI systems to learn from effectively
BHealthcare data is noisy, inconsistently structured, and often underrepresents diverse patient populations
CHealthcare data is too well-organized for AI to find meaningful patterns within it
DHealthcare organizations do not generate sufficient data for AI training purposes

3. A sepsis prediction AI trained on data from urban academic medical centers is deployed at a rural community hospital. The most significant governance concern is:

AThe AI will process data too slowly for the rural hospital's IT infrastructure
BThe model may perform poorly because the patient population differs from its training data
CRural hospitals do not have sufficient staffing to respond to AI-generated alerts
DThe AI system will require retraining every six months regardless of population

4. Which statement about general AI is most accurate?

AGeneral AI currently powers the most advanced diagnostic imaging systems in clinical use
BGeneral AI exists but is not yet approved for clinical deployment by regulatory authorities
CGeneral AI does not currently exist — all AI in clinical practice is narrow AI
DGeneral AI is the same as deep learning applied across multiple clinical domains simultaneously

5. Which AI application category is primarily concerned with analyzing unstructured clinical text to extract structured information?

ADiagnostic AI
BPredictive AI
CNatural Language Processing
DAdministrative and Operational AI
← Course Home Lesson 01 of 10 — AI Healthcare Quality & Safety Lesson 02 →